Method for matching satellite and UAV images for visual place recognition with cross-view color normalization
Year: 2025
Type: article
Abstract: The subject of this article is visual place recognition (VPR), specifically matching satellite images with images captured by unmanned aerial vehicles (UAVs). VPR is critical for autonomous UAV navigation, particularly in GPS-denied environments such as urban canyons or areas with significant infrastructure coverage where GNSS signals are unreliable. Despite its practical importance, accurately matching UAV images to satellite imagery remains challenging due to significant viewpoint, scale, illumination, and texture discrepancies. Traditional approaches that rely on handcrafted descriptors or classical local features often fail under such cross-view conditions. This study aims to design a robust visual place recognition method for matching UAV and satellite imagery, employing deep learning-based embeddings and advanced color normalization to improve reliability across cross-view scenarios. The tasks addressed in this article are: firstly, designing a YOLO-based method is designed for extracting global image embeddings, which utilizes YOLO’s multi-scale feature extraction capabilities to encode semantically significant landmarks in the scene. Second, a novel preprocessing technique based on aligning statistical color distributions between UAV and satellite images was developed and implemented to enhance their visual congruence. Finally, these components are integrated into a complete VPR system and evaluated for effectiveness using the challenging VPAIR dataset, emphasizing urban settings. The methods employed include deep learning techniques, particularly fine-tuning a YOLO11 neural network on a dataset specifically annotated for building segmentation. Statistical alignment techniques based on cumulative distribution functions (CDF) were used to standardize image appearances between the two distinct image domains. Conclusions. The experiments demonstrate significant improvements in UAV-to-satellite image matching performance using the proposed method. Fine-tuning YOLO11 specifically for building segmentation resulted in a robust embedding generation method that achieved high segmentation accuracy (F1-score of 0.722). The color preprocessing technique further improved the recognition performance, with Recall@1 reaching 19.5% for urban terrain within a localization radius of 3, substantially outperforming the traditional methods. This study provides an effective solution for UAV localization tasks, particularly in complex urban environments, highlighting the importance of integrated embedding extraction and domain-specific image preprocessing in cross-view visual place recognition.
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $107
OpenAlex ID: W4415547374